This question is about the intent to analyse taskwarrior data of people to improve automatic schedulers of tasks for individual taskwarrior users. This intend is accompanied with two desires. However I am not sure whether these two desires are realisable without compromise. The desires are:
- Users of taskwarrior are able to maintain privacy when they share their taskwarrior data. The user don't have to trust the data-analyst with their private data whilst still allowing their data to be used in data analysis.
- The second desire, is the ability for any arbitrary data-scientist to perform analysis on the data.
- The simplest conclusion I would draw is: one can't perform data analysis on data that is not seen/accessible.
- Yet, I thought
differential privacy might be a suitable tool to grant users privacy
whilst also enabling arbitrary datascientists to perform analysis on
that data. Though, I can't quite wrap my head around the privacy implications in scenarios where people enter identifying information along with a secret in a single "datapoint". For example, suppose it is a taboo to bake pancakes, and a person enters a task description
secretly bake pancakes *with my mother Joe at 4 O' clock on 1930 under the northern corner of this meadow at the blabla street 33 in Somecity*.This would contain a secret, e.g. baking pancakes, and it would identify the person because, suppose, there were only 1 person in this world with a mother named Joe that were on the northern corner of that meadow at that city.
I am not sure whether differential privacy is able to automatically obfuscate, with mathematical certainty whom the person was that baked pancakes, or what the person was doing at that time and place with that company. It is also not clear to me up to what extend this would require some intelligence to pre-process the data and set up the dataset that is parsed through a differential privacy software to yield an anonymised data set.
- I thought perhaps if the entire dataset is split into an encrypted sample- and an encrypted label dataset, that the relative patterns might be preserved, and I thought perhaps that would allow deep learning algorithms to train on that data and make predictions on new data that follows the same encryption.
Are there solutions that allow arbitrary data-scientists to perform data analysis on datasets whilst maintaining privacy of the persons whose data is in the datasets, without relying on the "intelligent" filtering of some person that preprocesses/prepares the data?